The main purpose of compaction is to keep reads fast by consolidating sstables together to avoid merging on read.
In a cluster with many tables we currently treat all pending compaction as equal. When in reality we may only be reading mainly from one of the tables.
Rather than FIFO we should prioritize access to the compactors based on read activity. SStables per read might be a good metric. Also, we would need to be sure to be fair to other tables over time. This would be a way to skew the work towards the tables who need compaction the most.
It might also be nice to offer a nodetool command to kill specific compaction jobs in progress that are not important under load.